Prove one provider call covers every present metric, absent metrics are
omitted from the prompt and never fabricated, an all-served cycle makes no
call (the per-metric hash gate still short-circuits), and the batch degrades
gracefully: an omitted metric or a provider error falls each affected card
back to its single-card path, invalid JSON retries once, and a no-provider
cycle serves each card's precomputed no-key result.
Add generateStatusBatchForUser: it prepares the seven specialised status
snapshots once, sends one prompt for a { perMetric } envelope covering only
the metrics with data, and fans each summary back into the same per-metric
cache rows the standalone generators write. One JSON-correction retry guards
the parse; any omitted metric or a failed batch degrades to that card's own
single-call path, so a partial response never goes dark. The nightly warm and
the forced warm route through the batch in production while the injected
per-card seam keeps the cron tests intact. Adds the status-batch prompt
scaffold composed from the shared assessment contract.
Apply the prepare -> run -> finalize split to blood-pressure, pulse, bmi,
mood, general, and medication-compliance status, matching weight-status.
Each public generator keeps its signature and one-call behaviour by driving
its prepared card through the shared single-card path; the snapshot build,
cache read, read-only miss, and content-hash gate are now reusable by a
batched caller without changing what each card persists.
Introduce a prepare -> run -> finalize contract for the per-metric status
cards so the snapshot build and the provider call can be driven separately.
The public generator keeps its signature and one-call-per-metric behaviour;
weight-status now resolves through the shared single-card path. Add the
status-batch AI budget entry sized for one combined seven-metric call.
Centralise per-surface AI budgets and right-size status maxTokens; add
seed support (pinned for reference surfaces, seedless for the briefing);
force JSON on Ollama and Anthropic; reroll only the daily-briefing prose
once a day for variety while findings stay stable; reuse the corrective
retry on the comprehensive path; extract shared prompt contracts as one
source of truth for grounding, tone, and the GLP-1 safety rule; reserve
the coach budget atomically before the call and reconcile burned tokens;
turn the local-AI private-host flag into a host allowlist.
`ALLOW_LOCAL_AI_PRIVATE_HOSTS` was a binary flag that, when `true`,
opened every private/internal host at once — including the cloud
metadata endpoint and any internal panel. Replace the binary check at
all three sites (write-time validation, local client, test resolver)
with a shared `isLocalAiHostAllowed` helper: a comma-separated host list
permits only those exact hostnames, while the legacy `true` still
permits any private host for backward compatibility. The safeFetch /
isPublicUrl floor is untouched; the allowlist only widens what the
operator deliberately permits. Docs and the SSRF-pin inventory test are
updated accordingly.
The daily token cap was read before the provider call and recorded only
after a successful reply, so concurrent requests could all read
`spent < cap` and all hit the provider, and an empty/sentinel reply
recorded zero spend despite burning upstream tokens.
Reserve the budget atomically before the call: a single upsert-increment
returns the new total, so concurrent requests serialise on the row and
cannot both pass the cap; over-cap refunds the reservation and surfaces
`coach.budget.exceeded` as an SSE error frame. Reconcile the reservation
to the actual token count immediately after the provider returns —
before any empty/sentinel short-circuit — so burned tokens are always
recorded. An integration test proves two concurrent requests cannot both
exceed the cap and that an empty reply still records its cost.
The per-worker in-process "last working provider" hint stays volatile
and is deliberately NOT folded into the durable Postgres health ledger.
The ledger already shares the load-bearing skip signal across workers;
the only uncoordinated piece is a soft reorder whose worst case is one
redundant first-hop per worker per hour. Promoting it to a per-request
ledger write would add hot-path write amplification for a negligible
saving. Documented inline so the trade-off is explicit.
The grounding, tone/banned-opener, GLP-1 dose-safety, metric-identifier
ban and forbidden-filler rules were independently encoded across the
comprehensive, status, Coach and narrative prompts and had drifted. Move
the canonical wording into `prompts/shared-contracts.ts` (locale-keyed
fragments) and compose each surface from it.
Where the surfaces had drifted, adopt the stricter wording: most notably
the GLP-1 dose-prescription contract — previously present only in the
comprehensive and Coach prompts — now reaches the status cards and the
period narrative too. A cross-surface coverage test asserts every
enforced fragment appears verbatim in every surface, in de and en.
The comprehensive generator did an ad-hoc JSON.parse with no corrective
retry and failed cold to `invalid-json` on a first-pass miss, while the
strict insight wrapper's `buildRetryCorrectionMessage` sat right next
door. Export that helper and reuse it: on a parse miss, append the
correction to the user prompt and re-run the chain once before declaring
failure. Unifies the JSON-robustness story across the two paths.
The content-hash gate re-stamps the prior text whenever the snapshot is
unchanged, so the "daily" briefing read byte-stable day-over-day. Decouple
phrasing freshness from data freshness for the dailyBriefing.paragraph
alone: on a hash-unchanged hit, re-roll just the paragraph once per
calendar day at a higher, seedless temperature, keeping every structured
finding verbatim from the cached payload.
A new `insightsBriefingRerollDate` column gates it to at most one extra
call per day, and a re-roll miss still stamps the day so a failing
provider cannot be re-driven on every page-open. The hash gate stays
fully intact for the status cards (reference assessments — monotony is
fine).
The measurement-reminder push already carries the reminder id in its
metadata, but the iOS allowlist stripped it before it reached the device.
Forward the key so a foreground or handled notification can post the
completion directly instead of routing through the capture deep link.
Add an opt-in `responseFormat: "json"` to CompletionParams that the
structured surfaces (comprehensive briefing + per-metric status cards)
set. On local (Ollama / OpenAI-compatible) it adds `format: "json"`; on
Anthropic it prefills the assistant turn with `{` and re-prepends the
brace to the reply, so the first emitted token is already inside a JSON
object. Both cut the first-pass JSON-failure rate that fence-stripping
otherwise has to absorb.
OpenAI already pins `response_format: json_object` and is unchanged. The
period narrative (plain text) and the Coach (streaming prose) leave the
flag unset, so their non-JSON paths are untouched.
Add an optional `seed` to CompletionParams and thread it onto the
OpenAI and local (Ollama / OpenAI-compatible) request bodies. The field
is dropped when unset, so existing callers are byte-identical.
Pin a stable seed for the status/reference cards (defaulted in the
status provider) and the period narrative, so a prompt change is
diff-able against a fixed baseline rather than masked by low-temperature
non-determinism. Anthropic's Messages API has no seed knob, so its
client ignores the field (documented inline). The seed is annotated on
the narrative wide-event.
Fix the target-edit-sheet invalidation that used a bare ["insights"] array
not matching the factory prefix the reads nest under (a silent no-op), and
migrate 12 further bare queryKey call sites onto the centralised factory.
Replace the scattered `?? 0.3` / `?? 1000` temperature and maxTokens
magic numbers across the status provider, Coach route, comprehensive
generator and period-narrative generator with a single typed
`ai-budgets.ts` map. Each entry documents the output contract that
fixes its token ceiling.
Right-size the per-metric status output from 1000 to 250 tokens: the
card renders one `{ "summary": "..." }` of 2-4 sentences (30-60 words),
so 1000 was an order of magnitude over the real ceiling.
The target editor invalidated a bare ["insights"] array on save and reset.
That literal sat outside the centralised factory, so the broad insights
fan-out it intended never matched the reads, which key through
insightsTargets/insightsRoot. Route the editor's reads and invalidations
through userThresholds(), insightsRoot(), insightsTargets() and analytics()
so a saved target actually refreshes every insights surface.
Migrate the other call sites the lint rule does not reach onto the factory:
the Coach conversation rail + thread reads, the per-metric personal-records
badge, the per-day measurement drill-down, the dashboard intake
compliance-chart fan-out, the per-medication side-effects list, the
onboarding auth invalidations, the notification-preferences read, and the
PWA persister's auth-me lookup.
Add the matching factory keys: coachConversations, coachConversation,
personalRecordsByMetric, measurementDrilldown, complianceChartInline and
medicationSideEffects. Each new key reuses the prefix its reads already use
so no invalidation shape drifts.
The dev-only SHA-256 key padding was gated on the absence of a production
marker, so a staging or preview container with NODE_ENV unset silently
derived a weak key the operator never recorded — rotating to a real key
later orphaned those rows (decrypt is fail-closed). Gate the padding on an
explicit NODE_ENV of development or test; every other case, including unset
NODE_ENV, throws the same hard error as in production.
coach_messages is append-only and grows forever — a daily Coach user
accrues encrypted conversation rows with no lifecycle, while every
sibling ledger (push_attempts, audit_logs, host_metrics) already prunes.
Add a daily pg-boss cleanup that hard-deletes rows past a configurable
window (COACH_MESSAGE_RETENTION_DAYS, default 365, 30-day floor so a
misconfig cannot drop live in-context history). Wire it into the
maintenance registrar with the dead-queue contract guard and a focused
helper test.
The demo seed's random-walk BP generator could drift below the
physiological floor and mint a systolic-0 row (iOS #33). Clamp the walk
to the same band the input validator enforces (systolic >= 40,
diastolic >= 20), reusing the VALUE_RANGES floors, so a fresh seed never
writes a sub-floor reading. Add a read-side selection guard on the blood
pressure series endpoint that excludes rows below those floors from both
the series and the implicit latest point, so a corrupt row that slipped
in before the floor was enforced can never surface as the latest
reading. Export VALUE_RANGES so both sites share one source of truth.
Add a coarseOnly option to buildTieredSeries that skips the raw 0–14d
read and the DAY (14–30d) / WEEK (30–90d) band reads, returning only the
MONTH / YEAR bands plus the anomaly envelope. The Coach snapshot's
coarse timeline tail consumes only month/year + anomalies, so the
near-term bands were read and discarded every turn; it now passes
coarseOnly. The briefing path is unchanged. Look the learn-guide prompt
example up by slug rather than by array index so reordering the catalog
cannot silently change the example link.
Drop steps from the signals-of-day set: Apple Health steps ingest as
many intraday stats-prefixed samples, so the newest raw row is a
partial-day fragment, not a daily total — same exclusion already applied
to per-night sleep. Add a glucose signal off the canonical stored value
(the model reads the snapshot, not display units) so the block matches
the priority comment that already claimed it; drop the dead sleep entry
from SIGNAL_PRIORITY. Track the raw extreme magnitude when keeping the
most extreme recent anomaly so a larger anomaly is no longer dropped by
comparing against an already-rounded value. Thread the injected clock
through trendSlope so the emerging-trend cutoff is deterministic under a
test clock. Rename the anomaly day-count to anomalyDaysAgo so the
snapshot hash staleness-tiers it instead of letting the raw count drive
the fingerprint. Cap the returned set at three to match the schema.
Wire the Coach snapshot's core-metric history onto the shared tiered
series so blood pressure, weight, and pulse carry a coarse tail the
in-window weekly fold could not produce — 90d-1y monthly and >1y yearly
buckets plus the preserved peak/trough envelope, so a spike from months
ago survives a coarse bucket. The tiered read skips the raw 0-14d band
the snapshot already holds, so no duplicate round-trip; the coarse tail
is bounded per band and sheds first under the prompt-budget degrader, so
token size holds.
Surface the active illness episodes and Rest Mode state to the Coach
prompt: the snapshot already carries the illness block, and the system
prompt now instructs the Coach to read an off-band vital as plausibly
illness-explained and avoid pushing a heavier cadence while the user is
in Rest Mode.
Add a static catalog of the public /learn guides and inject a compact
list into the system prompt. The Coach may reference at most one guide
by topic and URL when genuinely helpful, and only a slug the catalog
lists — a fabricated /learn URL is impossible by construction.
Enrich the Vorsorge cards with a discreet 7-day trend strip and a path
into the underlying readings.
- A quiet seven-bar strip of the reminder metric's last seven readings
sits under the metric context on both the dashboard summary card and
the full-grid cards, mirroring how a dashboard tile carries a small
trend. Neutral muted bars, no status colour. A free-text reminder or a
metric with fewer than two readings shows nothing, so card height stays
stable. The strip reads a bounded last-7 window per type through the
existing measurements endpoint, cached by type under the measurements
invalidation prefix.
- The card kebab gains a 'Show measurements' item for measurement-linked
reminders that opens /measurements pre-filtered to that type.
- The measurements list accepts a ?type=<MEASUREMENT_TYPE> deep link that
seeds the existing client-side type filter while leaving the rail
interactive; unknown tokens are dropped. The ?add= behaviour is
unchanged.
- New labels added across all six locales.
Render the new signalsOfDay list at the top of the daily-briefing card — a
present-focused row (headline + concrete nudge) sharing the tone bar, metric
icon, and sub-page routing of the key-finding row, with matching typography
(text-sm headline, text-xs muted body) so the card reads consistently with the
rest of the insight surfaces. Add the signalsTitle string across all six
locales.
Remove the "Ask the coach" action button and the guided-questions chip strip
from the overview hero band: the Coach stays the bottom-right drawer mounted by
the insights layout shell. The hero is now greeting + subtitle + baseline meta
+ the optional Health Score panel, and no longer reads the coach flags or the
launch context. Update the coach-cascade / disable-cascade fixtures and the
hero-strip slot tests to pin the affordances absent.
Compute a present-focused "signals of the day" block in extractFeatures from
the in-memory measurement set (no extra query): the freshest reading vs the
user's own 7d/30d windows, the normal-swing verdict, an emerging slope, and a
recent anomaly per salient metric, ranked by clinical priority and capped at
five. Add a signalsOfDay array to the dailyBriefing schema feeding the
briefing alongside the existing keyFindings cap, and rewrite the briefing
prompt (EN + DE) to lead with NOW: a warmer, motivating read that surfaces the
emerging signal and gives one concrete, doable nudge while still drawing on the
tiered history.
The signals block lands inside the compacted features payload, so it also
seeds insightsSnapshotHash — a fresh daily signal now changes the fingerprint
and forces the daily 04:30 warm to regenerate instead of re-stamping a
days-old narrative. Bump PROMPT_VERSION to 4.29.0.
Introduce buildTieredSeries(): a five-band progressive data-context helper —
0-14d raw daily, 14-30d DAY, 30-90d WEEK, 90d-1y MONTH, >1y YEAR — plus a
per-metric min/max-envelope anomaly list so a peak from months ago survives a
coarse bucket. Reuses probeRollupCoverage / ensureUserRollupsFresh with the
existing replace-fall-back-on-miss contract; readers never throw, so a
coverage gap simply yields an empty band. Caps each metric near 600 tokens to
stay within the existing context budget. Shared across the Coach snapshot and
the daily briefing.
The guided tour now exists only as the first-time auto-start after onboarding.
Remove every in-app trigger that re-opened it: the per-module "show this tour"
button on the dashboard, measurements, mood, medications, labs, illness, cycle,
achievements, insights, and coach surfaces; and the Settings replay card.
Delete the now-orphaned ModuleTourTrigger component, the tour-restart worker,
the healthlog:module-tour and healthlog:tour-restart event listeners, the
single-stop overlay path, and the force-launch session marker. The first-time
auto-launch gate and the tour overlay itself are unchanged. Prune the orphaned
i18n keys across all locales and update the affected tests.
buildTargetBands was a same-file alias of buildDashboardBands, which the
dead-code gate flags as a duplicate export. Keep the historic name only as
a re-export alias from the snapshot module, so callers and the parity test
are unchanged while bands.ts exposes a single builder.
Drop the health-score card's inert onAskCoach prop and its now-dead
pass-through from the hero strip, the buildDashboardBands re-export that no
caller reached through the snapshot module, an unused tour-progress status
type, and an orphan service-worker cache helper.
The dashboard recomputed the client band fallback on every render even when
the server snapshot supplied bands, and the module-list-prefs hook reparsed
its stored JSON into a fresh object each render, defeating downstream
memoisation. Wrap both in useMemo keyed on their real inputs.
The one-time disclaimer acknowledgment route shipped without an OpenAPI
entry, leaving a gap in the contract the iOS client consumes. Add the POST
path and regenerate the spec.
The proactive-nudge frequency gate only counted delivered push attempts. A
user with no push channel never wrote a delivered row, so the cap was inert
for them: every cron tick that hit a persistent trigger minted a fresh nudge
conversation in the rail.
Count a recently persisted nudge conversation toward the same rolling window
so the once-per-window contract holds regardless of push outcome. Add a
covering index on coach messages for the assistant-message scans the FAB poll
and the gate both run, and drop the unused providerType parameter from the
persistence helper now that the tag is fixed.
Add the GET /api/insights/coach/nudge-status unread signal and the POST
/api/insights/coach/seen open-stamp to the OpenAPI contract and regenerate
the spec.
Add POST /api/measurement-reminders/{id}/complete so a client can mark a
measurement reminder done on the server, rather than only dismissing the
notification locally. The route is the explicit user-action equivalent of
the cron auto-resolve: it routes through the one shared satisfaction
primitive, stamps lastSatisfiedAt, re-anchors the server-authoritative
nextDueAt, and fires no notification of its own.
Idempotent via the primitive's forward-only guard — completing an
already-completed or auto-satisfied reminder returns 200 with
completed=false so a repeat tap need not be special-cased. Owner-scoped;
a cross-user or tombstoned id 404s and never satisfies. No schema change:
lastSatisfiedAt already records the completion instant.
Extend the OpenAPI contract with the new path and the completion result
schema, and cover the route with unit tests (marks done, idempotent
no-op, owner-scoped 404, no double-notify).
Replace the per-page and per-chart disclaimer banners with a single
acknowledgment at the start of onboarding: HealthLog is a private tracking
tool, not a clinical assessment or diagnosis; talk to a doctor if worried.
- record the acknowledgment per user (disclaimer_acknowledged_at/_version,
migration 0175) mirroring the research-mode acknowledgment precedent
- gate the welcome "Get started" on the checkbox; expose the timestamp on
/api/auth/me so a returning user is never re-asked
- keep the full legal text reachable on the public privacy page
- add a wrench beside each module's add button linking to the module's own
settings page; rename every add button to a plain "hinzufügen"
- give each module a settings page: view (cards/list) + manual reorder; Labs
also gets a sort-order choice (by last reading) and biomarker CRUD/reorder
- collapse the Illness "abgeschlossen" list by default
- rebuild the Vorsorge card on the medication card's building blocks: a green
"Jetzt messen" action, the interval as a chip beside the name, and an
in-interval progress signal; drop the kebab "anpassen" and stray actions
- persist per-module view/order/sort client-side; add reverse-coverage-clean
strings across all six locales
The overview charted each trend but never offered a way into the metric's
detail page — the steps page in particular exists and is used, yet Insights
never linked to it. Give every Trends-row chart a routed detail target and a
discreet "view details" link below the caption, so a metric the row surfaces
(steps included, when the briefing flags it) drills into /insights/<slug>.
Add a discreet unread dot to the Coach FAB when the Coach has something
the user has not opened yet, and clear it when they open the Coach.
The signal is server-authoritative: a new coach_last_seen_at column on
users records the last time the Coach was opened, and the launcher's
status endpoint reports unread when the newest assistant message is
newer than that stamp — so the dot works even with no push channel and
clears across every device, web and iOS. Opening the Coach (FAB or the
/coach page) stamps the column via a new /api/insights/coach/seen
endpoint; the launcher keeps a local stamp only for an instant paint.
The dot is a calm cyan, not an alarming red, per the card-status rule.
The proactive Coach nudge dispatched a notification only — it never
wrote a CoachMessage, so it was invisible under the conversation list
and entirely silent for a user with no push channel configured.
Record the nudge as a real conversation before dispatching: create a
conversation and write the nudge body as the initial assistant message
(encrypted at rest like every other Coach reply), then fire the existing
notification + /coach deep link unchanged. A persistence failure is
caught locally so it can never swallow the notification, and the tick
summary tracks a separate persisted count alongside dispatched.
The reference-range classifier assumed bands[0] was always the
normal/target band, so any metric whose first band is abnormal was
placed wrong: a healthy BMI 22 read as just outside the reference and
an underweight BMI 17 read as inside; pulse pressure had the same flaw
(Narrow is its first band). The same classifier drives the in-app
four-state explainer, so both surfaces were affected.
Mark the population-normal band explicitly with a `normal: true` flag
on every metric, add `normalBandIndex`, and place values by distance
from the marked band rather than from index 0. `headlineBandText` now
renders the marked band, so BMI grounds against 18.5-24.9 kg/m2.
Exclude daily steps from the Coach grounding: it is fed the intra-day
mean of the step rows, which averages a multi-row day down and can
misplace it below the 8,000 floor. Grounding it on a per-day sum is the
correct fix and is left as a follow-up; until then drop the line rather
than ship a wrong placement.
Extend the classifier and grounding tests for BMI 17/22/27, a pulse
pressure case, the normal-band-not-at-index-0 contract, and that steps
is not grounded.
Surface the population reference bands behind the metrics a Coach turn
carries as a compact, citation-aware grounding block appended after the
SNAPSHOT. For each present metric the block states the headline band, its
source (blood pressure on the ESH 2023 line, US/ESC framings as context
only), and where the user's value sits via the shared four-state
placement contract.
Framed strictly as general guidance: the block opens and closes with a
not-a-diagnosis caveat, never overrides the existing refusal/dose/drug-
level safety discipline, and carries no commercial brand name. Glucose
respects the diabetes opt-in — the tighter ADA management goal only when
the user has declared it, otherwise the general non-diabetic band.
The block is deterministic and inspectable: built purely from the
reference-range backbone over the same scalars the snapshot already
surfaces, so it can never cite a value or band the snapshot does not
carry.
Wire absolute red-flag floors (BP >=180/120, SBP <90, glucose <70/<54/
>=250) into the illness notification path, escalating only on a confirmed
re-measured reading inside a 1h window, never diagnosing; asymptomatic ->
re-check/contact doctor, symptom-coupled -> emergency. Glucose gated behind
its module, BP core. Add an opt-in hasDiabetes pref (migration 0174) that
switches the glucose target resolver to the tighter ADA goal bands; never
auto-inferred. Resolver lives in the extracted targets builder.
Add absolute, guideline-backed safety floors that escalate a freshly logged
reading — but only after a second, confirming reading inside a one-hour
re-test window, so a single oscillometric spike or a one-off finger-stick can
never raise a false alarm:
- blood pressure 180/120 or higher (re-check, contact your doctor; symptom-
coupled lifts to emergency copy) and a symmetric low-systolic-under-90
cautionary tier;
- glucose under 70 (low) and under 54 (very low), and a sustained very-high
250-or-higher tier. The 200 DKA criterion deliberately never shows an
all-clear and never urgently escalates on its own, since a value alone
cannot rule euglycemic DKA in or out.
The escalation reuses the urgent SYSTEM_ALERT path with a 24h per-reason
dedup, never names a condition, and gates glucose behind the glucose module.
An optional transient `symptomsPresent` flag on the measurement write lifts a
confirmed breach from the contact-your-doctor copy to the emergency copy. The
measurement POST runs the check fire-and-forget so a notification failure
never fails the user's write.
Add a diabetes-aware glucose target resolver: when the user has opted in, the
fasting/post-prandial targets switch to the tighter ADA glycemic GOAL bands
(fasting 80–130, post-meal under 180 mg/dL) instead of the general
non-diabetic normal bands. A user threshold override always wins over both, so
a clinician-pinned individualised goal is never replaced by the population
band. The targets surface now reads through the resolver and labels the source
accordingly. The general band display is untouched.
Introduce an explicit, user-declared "I have diabetes / clinician glucose
targets" preference (`User.hasDiabetes`, migration 0174, defaulted false) plus
its GET/PATCH endpoint at `/api/auth/me/diabetes`. The flag selects which
glucose target band a reading is judged against; it is never inferred from a
value and never asserts a diagnosis. Mirrors the existing per-user-scalar
preference pattern: 60/min rate limit, Zod validation, audit-logged,
field-by-field write. OpenAPI regenerated for the new route.
Marc-André Bombeck